Method Development and Optimization of an Artificial Neural Network for Temperature Information Extraction from BOTDA Sensor System

We have shown in this article that with the change in the training data of artificial neural network generated to extract temperature information along the fiber, the output accuracy can be greatly improved and the sensitivity of the trained neural network to the noisy input can reduce a great extent. This change in the data includes adding a number of noisy Brillouin spectra with a signal-to-noise ratio of 25 to 30 to training dataset. It is worth noting that to date, the ideal input data without noise effects have been used for the purpose of training artificial neural networks to extract temperature or strain information along the fiber. It is also worth noting that in order to increase the output accuracy of the network, the network input spectrum needs a noise reduction process that the conventional method up to now is sequential averaging of the Brillouin spectra derived from the Brillouin optical time domain analyzer (BOTDA) sensor, which is reduces processing speed. In this article, we demonstrated that by adding noisy data to the networks training dataset, the sensitivity of the network to input noisy spectra also can greatly reduce, which results in a reduction in the number of successive averages of the signal to reduce input noise and consequently the processing speed increases. Also in this work, a comparison between neural network outputs that learned from the noisy input and a neural network that learned from the ideal input is presented and finally the superiority of the neural network that trained with noise effect is observable.

کلیدواژه ها:

BOTDA ، Brillouin Analyzer ، artificial neural network ، ANN ، Temperature ، strain ، optical fiber sensor